Department of
Mechanical, Automotive
and Aeronautical Engineering

KIM – Research

Our research develops and validates data-driven methods for the analysis, condition monitoring, and optimization of technical systems – and transfers them directly into industrial applications. The focus is on machine learning, deep learning, and statistical models for real industrial challenges.

At KIM, raw industrial data becomes reliable decision intelligence. Companies contribute their questions, equipment, and data – we provide the methodological foundation and experimental infrastructure. The result: deployment-ready AI solutions that shorten development cycles, stabilize processes, and reduce resource consumption.

Our Research Focus Areas

Our research combines state-of-the-art AI methods with engineering practice. The core focus areas are:

- Data-Driven Condition Forecasting: Development of ML models that detect wear and malfunctions at an early stage, enabling planned and predictable maintenance strategies.
- Predictive Quality: Application of statistical and AI-based methods to detect quality deviations in real time and minimize scrap.
- Adaptive Control & Self-Learning Agents: Use of reinforcement learning algorithms that autonomously optimize processes and inspection routines while adapting to changing conditions.
- Digital Twin & Simulation Analytics: Coupling of virtual twins with live data to run through scenarios, shorten development cycles, and reduce risks.
- Feature Engineering & Data Preparation: Scalable pipelines for data cleaning, synchronization, enrichment, and feature extraction – the foundation of reliable models.
- Explainable AI: Methods that make decision pathways more transparent, build trust, and meet regulatory requirements.

Real-World Application Areas

Quality Monitoring at End of Line: AI-assisted end-of-line inspection that evaluates sensor and image data captured throughout the process to immediately detect deviations and minimize scrap.
Anomaly Detection in Safety-Critical Systems: Analysis of image, signal, and telemetry data for the early detection of unusual patterns and risks, ensuring the highest levels of operational safety.
Trajectory Planning for Autonomous Drones: Reinforcement learning-based route optimization with flexibly adjustable constraints – for efficient, safe, and energy-saving flight paths.
Usage Scenario & User Profile Analysis: Evaluation of field data to identify real-world usage patterns and user behavior. The insights gained enable more targeted test plans and customer-oriented, needs-based product design.
Meta-Modeling & Surrogate Models: Development of fast, data-driven substitute models based on diverse simulations. These support early design decisions, reduce computational effort in the development process, and accelerate variant studies.

Knife Blades

To objectively evaluate fine-ground knife blade surfaces, we apply image-based analysis and AI-assisted classification. Geometric features are automatically extracted and correlated with manufacturing parameters – the results provide direct feedback for process optimization.

Anomaly Detection

Fine scratches, cracks, or structural deviations on metal surfaces are often the first sign of a manufacturing defect. Using statistical methods and AI-based image analysis, our system detects such anomalies automatically and reliably – before they lead to scrap.

AI-Driven Drone Swarm Coordination for Disaster Response

Autonomous drone fleets (UAVs) are deployed for reconnaissance in flood-affected areas. Using Deep Reinforcement Learning and geospatial data fusion, the agents learn dynamic navigation strategies for the autonomous identification of critical infrastructure and damaged zones.

Process Monitoring in Microalgae Cultivation

Microscopy images of individual algae cells are automatically segmented, and features describing shape, color, texture, and intensity distribution are extracted from each cell. Unsupervised clustering methods then identify distinct growth phases and cell states – without any manual labeling. The result is an image-based sensor that complements conventional measurement techniques and provides the foundation for predictive growth models.

Stress Prediction Using Physics-Informed Graph Neural Networks

FEM simulations deliver accurate stress fields but are computationally expensive. Graph Neural Networks (GNNs) directly exploit the mesh structure of FEM models, enabling much faster predictions once trained. Physics-informed PINN terms in the loss function specifically improve the prediction of safety-critical notch stresses at weld transitions.

Cooperation Partners

Technical Equipment

Deep Learning Workstation with 2x NVIDIA RTX A5500 GPUs uand 128GB DDR4-RAM
Deep Learning Workstation with 2x NVIDIA RTX 6000 Ada Lovelace GPUS, 1x NVIDIA RTX 6000 Blackwell Pro Max-Q GPU and 192GB DDR4-RAM
Access to the BayernKI-Cluster at LRZ, including 320x NVIDIA H100 GPUs
Condition Monitoring & Predictive Quality Test Bench with multiple cameras (Fujifilm X100VI, GoPro HERO12, and industrial CCD cameras) for the acquisition of complex workpiece surfaces